Pull scheduler fixes from Ingo Molnar:
"Misc fixes all over the place:
- Fix NUMA over-balancing between lightly loaded nodes. This is
fallout of the big load-balancer rewrite.
- Fix the NOHZ remote loadavg update logic, which fixes anomalies
like reported 150 loadavg on mostly idle CPUs.
- Fix XFS performance/scalability
- Fix throttled groups unbound task-execution bug
- Fix PSI procfs boundary condition
- Fix the cpu.uclamp.{min,max} cgroup configuration write checks
- Fix DocBook annotations
- Fix RCU annotations
- Fix overly CPU-intensive housekeeper CPU logic loop on large CPU
counts"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/fair: Fix kernel-doc warning in attach_entity_load_avg()
sched/core: Annotate curr pointer in rq with __rcu
sched/psi: Fix OOB write when writing 0 bytes to PSI files
sched/fair: Allow a per-CPU kthread waking a task to stack on the same CPU, to fix XFS performance regression
sched/fair: Prevent unlimited runtime on throttled group
sched/nohz: Optimize get_nohz_timer_target()
sched/uclamp: Reject negative values in cpu_uclamp_write()
sched/fair: Allow a small load imbalance between low utilisation SD_NUMA domains
timers/nohz: Update NOHZ load in remote tick
sched/core: Don't skip remote tick for idle CPUs
Pull power management fixes from Rafael Wysocki:
"Fix three issues related to the handling of wakeup events signaled
through the ACPI SCI while suspended to idle (Rafael Wysocki) and
unexport an internal cpufreq variable (Yangtao Li)"
* tag 'pm-5.6-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
ACPI: PM: s2idle: Prevent spurious SCIs from waking up the system
ACPICA: Introduce acpi_any_gpe_status_set()
ACPI: PM: s2idle: Avoid possible race related to the EC GPE
ACPI: EC: Fix flushing of pending work
cpufreq: Make cpufreq_global_kobject static
Because cpuidle is the only user of the effective constraint coming
from the CPU latency QoS, add #ifdef CONFIG_CPU_IDLE around that code
to avoid building it unnecessarily.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Update the file information comments in include/linux/pm_qos.h
and kernel/power/qos.c by adding titles along with copyright and
authors information to them and changing the qos.c description to
better reflect its contents (outdated information is dropped from
it in particular).
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Drop the PM QoS classes enum including PM_QOS_CPU_DMA_LATENCY,
drop the wrappers around pm_qos_request(), pm_qos_request_active(),
and pm_qos_add/update/remove_request() introduced previously, rename
these functions, respectively, to cpu_latency_qos_limit(),
cpu_latency_qos_request_active(), and
cpu_latency_qos_add/update/remove_request(), and update their
kerneldoc comments. [While at it, drop some useless comments from
these functions.]
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Qian Cai reported that the WARN_ON() in the x86/msi affinity setting code,
which catches cases where the affinity setting is not done on the CPU which
is the current target of the interrupt, triggers during CPU hotplug stress
testing.
It turns out that the warning which was added with the commit addressing
the MSI affinity race unearthed yet another long standing bug.
If user space writes a bogus affinity mask, i.e. it contains no online CPUs,
then it calls irq_select_affinity_usr(). This was introduced for ALPHA in
eee45269b0 ("[PATCH] Alpha: convert to generic irq framework (generic part)")
and subsequently made available for all architectures in
1840475676 ("genirq: Expose default irq affinity mask (take 3)")
which introduced the circumvention of the affinity setting restrictions for
interrupt which cannot be moved in process context.
The whole exercise is bogus in various aspects:
1) If the interrupt is already started up then there is absolutely
no point to honour a bogus interrupt affinity setting from user
space. The interrupt is already assigned to an online CPU and it
does not make any sense to reassign it to some other randomly
chosen online CPU.
2) If the interupt is not yet started up then there is no point
either. A subsequent startup of the interrupt will invoke
irq_setup_affinity() anyway which will chose a valid target CPU.
So the only correct solution is to just return -EINVAL in case user space
wrote an affinity mask which does not contain any online CPUs, except for
ALPHA which has it's own magic sauce for this.
Fixes: 1840475676 ("genirq: Expose default irq affinity mask (take 3)")
Reported-by: Qian Cai <cai@lca.pw>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Qian Cai <cai@lca.pw>
Link: https://lkml.kernel.org/r/878sl8xdbm.fsf@nanos.tec.linutronix.de
Change the return type of pm_qos_request() to be the same as the
one of pm_qos_read_value() called by it internally and stop exporting
it to modules (because its only caller, cpuidle, is not modular).
Also move the pm_qos_read_value() header away from the CPU latency
QoS API function headers in pm_qos.h (because it technically does
not belong to that API).
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Modify the definitions of the CPU latency QoS trace events to take
one argument (since PM_QOS_CPU_DMA_LATENCY is always passed as the
pm_qos_class argument to them) and update the documentation of them
accordingly (while at it, make it explicitly mention CPU latency QoS
and relocate it after the device PM QoS trace events documentation).
The names and output format of the trace events do not change to
preserve user space compatibility.
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
First, rename PM_QOS_CPU_DMA_LAT_DEFAULT_VALUE to
PM_QOS_CPU_LATENCY_DEFAULT_VALUE and update all of the code
referring to it accordingly.
Next, rename cpu_dma_constraints to cpu_latency_constraints, move
the definition of it closer to the functions referring to it and
update all of them accordingly. [While at it, add a comment to mark
the start of the code related to the CPU latency QoS.]
Finally, rename the pm_qos_power_*() family of functions and
pm_qos_power_fops to cpu_latency_qos_*() and cpu_latency_qos_fops,
respectively, and update the definition of cpu_latency_qos_miscdev.
[While at it, update the miscdev interface code start comment.]
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Notice that pm_qos_remove_notifier() is not used at all and the only
caller of pm_qos_add_notifier() is the cpuidle core, which only needs
the PM_QOS_CPU_DMA_LATENCY notifier to invoke wake_up_all_idle_cpus()
upon changes of the PM_QOS_CPU_DMA_LATENCY target value.
First, to ensure that wake_up_all_idle_cpus() will be called
whenever the PM_QOS_CPU_DMA_LATENCY target value changes, modify the
pm_qos_add/update/remove_request() family of functions to check if
the effective constraint for the PM_QOS_CPU_DMA_LATENCY has changed
and call wake_up_all_idle_cpus() directly in that case.
Next, drop the PM_QOS_CPU_DMA_LATENCY notifier from cpuidle as it is
not necessary any more.
Finally, drop both pm_qos_add_notifier() and pm_qos_remove_notifier(),
as they have no callers now, along with cpu_dma_lat_notifier which is
only used by them.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
First, change the definition of struct pm_qos_request so that it
contains a struct pm_qos_constraints pointer (called "qos") instead
of a PM QoS class number (in preparation for dropping the PM QoS
classes concept altogether going forward) and move its definition
(along with the definition of struct pm_qos_flags_request that does
not change) after the definition of struct pm_qos_constraints.
Next, drop the definition of struct pm_qos_object and the null_pm_qos
and cpu_dma_pm_qos variables of that type along with pm_qos_array[]
holding pointers to them and change the code to refer to the
pm_qos_constraints structure directly or to use the new qos pointer
in struct pm_qos_request for that instead of going through
pm_qos_array[] to access it. Also update kerneldoc comments that
mention pm_qos_class to refer to PM_QOS_CPU_DMA_LATENCY directly
instead.
Finally, drop register_pm_qos_misc(), introduce cpu_latency_qos_miscdev
(with the name field set to "cpu_dma_latency") to implement the
CPU latency QoS interface in /dev/ and register it directly from
pm_qos_power_init().
After these changes the notion of PM QoS classes remains only in the
API (in the form of redundant function parameters that are ignored)
and in the definitions of PM QoS trace events.
While at it, some redundant local variables are dropped etc.
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Reorder the code to avoid using extra function header declarations
for the pm_qos_power_*() family of functions and drop those
declarations.
Also clean up the internals of those functions to consolidate checks,
avoid using redundant local variables and similar.
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
After commit c3082a674f ("PM: QoS: Get rid of unused flags") the
only global PM QoS class in use is PM_QOS_CPU_DMA_LATENCY, so it
does not really make sense to iterate over global QoS classes
anywhere, since there is only one.
Remove iterations over global QoS classes from the code and use
PM_QOS_CPU_DMA_LATENCY as the target class directly where needed.
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Move the definition of pm_qos_read_value() before the one of
pm_qos_get_value() and add a kerneldoc comment to it (as it is
not static).
Also replace the BUG() in pm_qos_get_value() with WARN() (to
prevent the kernel from crashing if an unknown PM QoS type is
used by mistake) and drop the comment next to it that is not
necessary any more.
Additionally, drop the unnecessary inline modifier from the header
of pm_qos_set_value().
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
Clean up the pm_qos_update_target() function:
* Update its kerneldoc comment.
* Drop the redundant ret local variable from it.
* Reorder definitions of local variables in it.
* Update a comment in it.
Also update the kerneldoc comment of pm_qos_update_flags() (e.g.
notifiers are not called by it any more) and add one emtpy line
to its body (for more visual clarity).
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
The PM_QOS_SUM QoS type is not used, so drop it along with the
code referring to it in pm_qos_get_value() and the related local
variables in there.
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
The pm_qos_update_request_timeout() function is not called from
anywhere, so drop it along with the work member in struct
pm_qos_request needed by it.
Also drop the useless pm_qos_update_request_timeout trace event
that is only triggered by that function (so it never triggers at
all) and update the trace events documentation accordingly.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Tested-by: Amit Kucheria <amit.kucheria@linaro.org>
This adds support for creating a process in a different cgroup than its
parent. Callers can limit and account processes and threads right from
the moment they are spawned:
- A service manager can directly spawn new services into dedicated
cgroups.
- A process can be directly created in a frozen cgroup and will be
frozen as well.
- The initial accounting jitter experienced by process supervisors and
daemons is eliminated with this.
- Threaded applications or even thread implementations can choose to
create a specific cgroup layout where each thread is spawned
directly into a dedicated cgroup.
This feature is limited to the unified hierarchy. Callers need to pass
a directory file descriptor for the target cgroup. The caller can
choose to pass an O_PATH file descriptor. All usual migration
restrictions apply, i.e. there can be no processes in inner nodes. In
general, creating a process directly in a target cgroup adheres to all
migration restrictions.
One of the biggest advantages of this feature is that CLONE_INTO_GROUP does
not need to grab the write side of the cgroup cgroup_threadgroup_rwsem.
This global lock makes moving tasks/threads around super expensive. With
clone3() this lock is avoided.
Cc: Tejun Heo <tj@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: cgroups@vger.kernel.org
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The core codepaths to check whether a process can be attached to a
cgroup are the same for threads and thread-group leaders. Only a small
piece of code verifying that source and destination cgroup are in the
same domain differentiates the thread permission checking from
thread-group leader permission checking.
Since cgroup_migrate_vet_dst() only matters cgroup2 - it is a noop on
cgroup1 - we can move it out of cgroup_attach_task().
All checks can now be consolidated into a new helper
cgroup_attach_permissions() callable from both cgroup_procs_write() and
cgroup_threads_write().
Cc: Tejun Heo <tj@kernel.org>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: cgroups@vger.kernel.org
Acked-by: Michal Koutný <mkoutny@suse.com>
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Convert cpuset_hotplug_workfn() into synchronous call for cpu hotplug
path. For memory hotplug path it still gets queued as a work item.
Since cpuset_hotplug_workfn() can be made synchronous for cpu hotplug
path, it is not required to wait for cpuset hotplug while thawing
processes.
Signed-off-by: Prateek Sood <prsood@codeaurora.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
list_for_each_entry_rcu has built-in RCU and lock checking.
Pass cond argument to list_for_each_entry_rcu() to silence
false lockdep warning when CONFIG_PROVE_RCU_LIST is enabled
by default.
Even though the function css_next_child() already checks if
cgroup_mutex or rcu_read_lock() is held using
cgroup_assert_mutex_or_rcu_locked(), there is a need to pass
cond to list_for_each_entry_rcu() to avoid false positive
lockdep warning.
Signed-off-by: Madhuparna Bhowmik <madhuparnabhowmik10@gmail.com>
Acked-by: Michal Koutný <mkoutny@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
css_task_iter stores pointer to head of each iterable list, this dates
back to commit 0f0a2b4fa6 ("cgroup: reorganize css_task_iter") when we
did not store cur_cset. Let us utilize list heads directly in cur_cset
and streamline css_task_iter_advance_css_set a bit. This is no
intentional function change.
Signed-off-by: Michal Koutný <mkoutny@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
PF_EXITING is set earlier than actual removal from css_set when a task
is exitting. This can confuse cgroup.procs readers who see no PF_EXITING
tasks, however, rmdir is checking against css_set membership so it can
transitionally fail with EBUSY.
Fix this by listing tasks that weren't unlinked from css_set active
lists.
It may happen that other users of the task iterator (without
CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This
is equal to the state before commit c03cd7738a ("cgroup: Include dying
leaders with live threads in PROCS iterations") but it may be reviewed
later.
Reported-by: Suren Baghdasaryan <surenb@google.com>
Fixes: c03cd7738a ("cgroup: Include dying leaders with live threads in PROCS iterations")
Signed-off-by: Michal Koutný <mkoutny@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
If seq_file .next fuction does not change position index,
read after some lseek can generate unexpected output:
1) dd bs=1 skip output of each 2nd elements
$ dd if=/sys/fs/cgroup/cgroup.procs bs=8 count=1
2
3
4
5
1+0 records in
1+0 records out
8 bytes copied, 0,000267297 s, 29,9 kB/s
[test@localhost ~]$ dd if=/sys/fs/cgroup/cgroup.procs bs=1 count=8
2
4 <<< NB! 3 was skipped
6 <<< ... and 5 too
8 <<< ... and 7
8+0 records in
8+0 records out
8 bytes copied, 5,2123e-05 s, 153 kB/s
This happen because __cgroup_procs_start() makes an extra
extra cgroup_procs_next() call
2) read after lseek beyond end of file generates whole last line.
3) read after lseek into middle of last line generates
expected rest of last line and unexpected whole line once again.
Additionally patch removes an extra position index changes in
__cgroup_procs_start()
Cc: stable@vger.kernel.orghttps://bugzilla.kernel.org/show_bug.cgi?id=206283
Signed-off-by: Vasily Averin <vvs@virtuozzo.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
if seq_file .next fuction does not change position index,
read after some lseek can generate unexpected output.
# mount | grep cgroup
# dd if=/mnt/cgroup.procs bs=1 # normal output
...
1294
1295
1296
1304
1382
584+0 records in
584+0 records out
584 bytes copied
dd: /mnt/cgroup.procs: cannot skip to specified offset
83 <<< generates end of last line
1383 <<< ... and whole last line once again
0+1 records in
0+1 records out
8 bytes copied
dd: /mnt/cgroup.procs: cannot skip to specified offset
1386 <<< generates last line anyway
0+1 records in
0+1 records out
5 bytes copied
https://bugzilla.kernel.org/show_bug.cgi?id=206283
Signed-off-by: Vasily Averin <vvs@virtuozzo.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
After commit c3082a674f ("PM: QoS: Get rid of unused flags") the
only global PM QoS class in use is PM_QOS_CPU_DMA_LATENCY and the
existing PM QoS debugfs interface has become overly complicated (as
it takes other potentially possible PM QoS classes that are not there
any more into account). It is also not particularly useful (the
"type" of the PM_QOS_CPU_DMA_LATENCY is known, its aggregate value
can be read from /dev/cpu_dma_latency and the number of requests in
the queue does not really matter) and there are no known users
depending on it. Moreover, there are dedicated trace events that
can be used for tracking PM QoS usage with much higher precision.
For these reasons, drop the PM QoS debugfs interface altogether.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Pull tracing fixes from Steven Rostedt:
"Various fixes:
- Fix an uninitialized variable
- Fix compile bug to bootconfig userspace tool (in tools directory)
- Suppress some error messages of bootconfig userspace tool
- Remove unneded CONFIG_LIBXBC from bootconfig
- Allocate bootconfig xbc_nodes dynamically. To ease complaints about
taking up static memory at boot up
- Use of parse_args() to parse bootconfig instead of strstr() usage
Prevents issues of double quotes containing the interested string
- Fix missing ring_buffer_nest_end() on synthetic event error path
- Return zero not -EINVAL on soft disabled synthetic event (soft
disabling must be the same as hard disabling, which returns zero)
- Consolidate synthetic event code (remove duplicate code)"
* tag 'trace-v5.6-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Consolidate trace() functions
tracing: Don't return -EINVAL when tracing soft disabled synth events
tracing: Add missing nest end to synth_event_trace_start() error case
tools/bootconfig: Suppress non-error messages
bootconfig: Allocate xbc_nodes array dynamically
bootconfig: Use parse_args() to find bootconfig and '--'
tracing/kprobe: Fix uninitialized variable bug
bootconfig: Remove unneeded CONFIG_LIBXBC
tools/bootconfig: Fix wrong __VA_ARGS__ usage
The low level index is the index in the underlying hardware buffer of
the most recently captured taken branch which is always saved in
branch_entries[0]. It is very useful for reconstructing the call stack.
For example, in Intel LBR call stack mode, the depth of reconstructed
LBR call stack limits to the number of LBR registers. With the low level
index information, perf tool may stitch the stacks of two samples. The
reconstructed LBR call stack can break the HW limitation.
Add a new branch sample type to retrieve low level index of raw branch
records. The low level index is between -1 (unknown) and max depth which
can be retrieved in /sys/devices/cpu/caps/branches.
Only when the new branch sample type is set, the low level index
information is dumped into the PERF_SAMPLE_BRANCH_STACK output.
Perf tool should check the attr.branch_sample_type, and apply the
corresponding format for PERF_SAMPLE_BRANCH_STACK samples.
Otherwise, some user case may be broken. For example, users may parse a
perf.data, which include the new branch sample type, with an old version
perf tool (without the check). Users probably get incorrect information
without any warning.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200127165355.27495-2-kan.liang@linux.intel.com
The filesystem freezer uses percpu-rwsem in a way that is effectively
write_non_owner() and achieves this with a few horrible hacks that
rely on the rwsem (!percpu) implementation.
When PREEMPT_RT replaces the rwsem implementation with a PI aware
variant this comes apart.
Remove the embedded rwsem and implement it using a waitqueue and an
atomic_t.
- make readers_block an atomic, and use it, with the waitqueue
for a blocking test-and-set write-side.
- have the read-side wait for the 'lock' state to clear.
Have the waiters use FIFO queueing and mark them (reader/writer) with
a new WQ_FLAG. Use a custom wake_function to wake either a single
writer or all readers until a writer.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Davidlohr Bueso <dbueso@suse.de>
Acked-by: Will Deacon <will@kernel.org>
Acked-by: Waiman Long <longman@redhat.com>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lkml.kernel.org/r/20200204092403.GB14879@hirez.programming.kicks-ass.net
Once a lock class is zapped, all the lock chains that include the zapped
class are essentially useless. The lock_chain structure itself can be
reused, but not the corresponding chain_hlocks[] entries. Over time,
we will run out of chain_hlocks entries while there are still plenty
of other lockdep array entries available.
To fix this imbalance, we have to make chain_hlocks entries reusable
just like the others. As the freed chain_hlocks entries are in blocks of
various lengths. A simple bitmap like the one used in the other reusable
lockdep arrays isn't applicable. Instead the chain_hlocks entries are
put into bucketed lists (MAX_CHAIN_BUCKETS) of chain blocks. Bucket 0
is the variable size bucket which houses chain blocks of size larger than
MAX_CHAIN_BUCKETS sorted in decreasing size order. Initially, the whole
array is in one chain block (the primordial chain block) in bucket 0.
The minimum size of a chain block is 2 chain_hlocks entries. That will
be the minimum allocation size. In other word, allocation requests
for one chain_hlocks entry will cause 2-entry block to be returned and
hence 1 entry will be wasted.
Allocation requests for the chain_hlocks are fulfilled first by looking
for chain block of matching size. If not found, the first chain block
from bucket[0] (the largest one) is split. That can cause hlock entries
fragmentation and reduce allocation efficiency if a chain block of size >
MAX_CHAIN_BUCKETS is ever zapped and put back to after the primordial
chain block. So the MAX_CHAIN_BUCKETS must be large enough that this
should seldom happen.
By reusing the chain_hlocks entries, we are able to handle workloads
that add and zap a lot of lock classes without the risk of running out
of chain_hlocks entries as long as the total number of outstanding lock
classes at any time remain within a reasonable limit.
Two new tracking counters, nr_free_chain_hlocks & nr_large_chain_blocks,
are added to track the total number of chain_hlocks entries in the
free bucketed lists and the number of large chain blocks in buckets[0]
respectively. The nr_free_chain_hlocks replaces nr_chain_hlocks.
The nr_large_chain_blocks counter enables to see if we should increase
the number of buckets (MAX_CHAIN_BUCKETS) available so as to avoid to
avoid the fragmentation problem in bucket[0].
An internal nfsd test that ran for more than an hour and kept on
loading and unloading kernel modules could cause the following message
to be displayed.
[ 4318.443670] BUG: MAX_LOCKDEP_CHAIN_HLOCKS too low!
The patched kernel was able to complete the test with a lot of free
chain_hlocks entries to spare:
# cat /proc/lockdep_stats
:
dependency chains: 18867 [max: 65536]
dependency chain hlocks: 74926 [max: 327680]
dependency chain hlocks lost: 0
:
zapped classes: 1541
zapped lock chains: 56765
large chain blocks: 1
By changing MAX_CHAIN_BUCKETS to 3 and add a counter for the size of the
largest chain block. The system still worked and We got the following
lockdep_stats data:
dependency chains: 18601 [max: 65536]
dependency chain hlocks used: 73133 [max: 327680]
dependency chain hlocks lost: 0
:
zapped classes: 1541
zapped lock chains: 56702
large chain blocks: 45165
large chain block size: 20165
By running the test again, I was indeed able to cause chain_hlocks
entries to get lost:
dependency chain hlocks used: 74806 [max: 327680]
dependency chain hlocks lost: 575
:
large chain blocks: 48737
large chain block size: 7
Due to the fragmentation, it is possible that the
"MAX_LOCKDEP_CHAIN_HLOCKS too low!" error can happen even if a lot of
of chain_hlocks entries appear to be free.
Fortunately, a MAX_CHAIN_BUCKETS value of 16 should be big enough that
few variable sized chain blocks, other than the initial one, should
ever be present in bucket 0.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200206152408.24165-7-longman@redhat.com
If a lock chain contains a class that is zapped, the whole lock chain is
likely to be invalid. If the zapped class is at the end of the chain,
the partial chain without the zapped class should have been stored
already as the current code will store all its predecessor chains. If
the zapped class is somewhere in the middle, there is no guarantee that
the partial chain will actually happen. It may just clutter up the hash
and make searching slower. I would rather prefer storing the chain only
when it actually happens.
So just dump the corresponding chain_hlocks entries for now. A latter
patch will try to reuse the freed chain_hlocks entries.
This patch also changes the type of nr_chain_hlocks to unsigned integer
to be consistent with the other counters.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200206152408.24165-5-longman@redhat.com
The whole point of the lockdep dynamic key patch is to allow unused
locks to be removed from the lockdep data buffers so that existing
buffer space can be reused. However, there is no way to find out how
many unused locks are zapped and so we don't know if the zapping process
is working properly.
Add a new nr_zapped_classes counter to track that and show it in
/proc/lockdep_stats.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200206152408.24165-4-longman@redhat.com
Currently, the irq_context field of a lock chains displayed in
/proc/lockdep_chains is just a number. It is likely that many people
may not know what a non-zero number means. To make the information more
useful, print the actual irq names ("softirq" and "hardirq") instead.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200206152408.24165-3-longman@redhat.com
There are currently three counters to track the IRQ context of a lock
chain - nr_hardirq_chains, nr_softirq_chains and nr_process_chains.
They are incremented when a new lock chain is added, but they are
not decremented when a lock chain is removed. That causes some of the
statistic counts reported by /proc/lockdep_stats to be incorrect.
IRQ
Fix that by decrementing the right counter when a lock chain is removed.
Since inc_chains() no longer accesses hardirq_context and softirq_context
directly, it is moved out from the CONFIG_TRACE_IRQFLAGS conditional
compilation block.
Fixes: a0b0fd53e1 ("locking/lockdep: Free lock classes that are no longer in use")
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200206152408.24165-2-longman@redhat.com
